1 /*****************************************************************************
2 * ratecontrol.c: h264 encoder library (Rate Control)
3 *****************************************************************************
4 * Copyright (C) 2005-2008 x264 project
6 * Authors: Loren Merritt <lorenm@u.washington.edu>
7 * Michael Niedermayer <michaelni@gmx.at>
8 * Gabriel Bouvigne <gabriel.bouvigne@joost.com>
9 * Fiona Glaser <fiona@x264.com>
10 * Måns Rullgård <mru@mru.ath.cx>
12 * This program is free software; you can redistribute it and/or modify
13 * it under the terms of the GNU General Public License as published by
14 * the Free Software Foundation; either version 2 of the License, or
15 * (at your option) any later version.
17 * This program is distributed in the hope that it will be useful,
18 * but WITHOUT ANY WARRANTY; without even the implied warranty of
19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
20 * GNU General Public License for more details.
22 * You should have received a copy of the GNU General Public License
23 * along with this program; if not, write to the Free Software
24 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111, USA.
25 *****************************************************************************/
27 #define _ISOC99_SOURCE
28 #undef NDEBUG // always check asserts, the speed effect is far too small to disable them
31 #include "common/common.h"
32 #include "common/cpu.h"
33 #include "ratecontrol.h"
45 uint64_t expected_bits; /*total expected bits up to the current frame (current one excluded)*/
52 float blurred_complexity;
55 int16_t i_weight_denom;
60 } ratecontrol_entry_t;
70 struct x264_ratecontrol_t
79 double rate_tolerance;
81 int nmb; /* number of macroblocks in a frame */
85 ratecontrol_entry_t *rce;
86 int qp; /* qp for current frame */
87 int qpm; /* qp for current macroblock */
88 float f_qpm; /* qp for current macroblock: precise float for AQ */
89 float qpa_rc; /* average of macroblocks' qp before aq */
90 float qpa_aq; /* average of macroblocks' qp after aq */
91 float qp_novbv; /* QP for the current frame if 1-pass VBV was disabled. */
96 double buffer_fill_final; /* real buffer as of the last finished frame */
97 double buffer_fill; /* planned buffer, if all in-progress frames hit their bit budget */
98 double buffer_rate; /* # of bits added to buffer_fill after each frame */
99 double vbv_max_rate; /* # of bits added to buffer_fill per second */
100 predictor_t *pred; /* predict frame size from satd */
101 int single_frame_vbv;
102 double rate_factor_max_increment; /* Don't allow RF above (CRF + this value). */
107 double cplxr_sum; /* sum of bits*qscale/rceq */
108 double expected_bits_sum; /* sum of qscale2bits after rceq, ratefactor, and overflow, only includes finished frames */
109 double wanted_bits_window; /* target bitrate * window */
111 double short_term_cplxsum;
112 double short_term_cplxcount;
113 double rate_factor_constant;
118 FILE *p_stat_file_out;
119 char *psz_stat_file_tmpname;
120 FILE *p_mbtree_stat_file_out;
121 char *psz_mbtree_stat_file_tmpname;
122 char *psz_mbtree_stat_file_name;
123 FILE *p_mbtree_stat_file_in;
125 int num_entries; /* number of ratecontrol_entry_ts */
126 ratecontrol_entry_t *entry; /* FIXME: copy needed data and free this once init is done */
128 double last_qscale_for[5]; /* last qscale for a specific pict type, used for max_diff & ipb factor stuff */
129 int last_non_b_pict_type;
130 double accum_p_qp; /* for determining I-frame quant */
132 double last_accum_p_norm;
133 double lmin[5]; /* min qscale by frame type */
135 double lstep; /* max change (multiply) in qscale per frame */
136 uint16_t *qp_buffer[2]; /* Global buffers for converting MB-tree quantizer data. */
137 int qpbuf_pos; /* In order to handle pyramid reordering, QP buffer acts as a stack.
138 * This value is the current position (0 or 1). */
141 float frame_size_estimated; /* Access to this variable must be atomic: double is
142 * not atomic on all arches we care about */
143 double frame_size_maximum; /* Maximum frame size due to MinCR */
144 double frame_size_planned;
145 double slice_size_planned;
146 double max_frame_error;
147 predictor_t (*row_pred)[2];
148 predictor_t row_preds[5][2];
149 predictor_t *pred_b_from_p; /* predict B-frame size from P-frame satd */
150 int bframes; /* # consecutive B-frames before this P-frame */
151 int bframe_bits; /* total cost of those frames */
155 x264_zone_t *prev_zone;
158 int initial_cpb_removal_delay;
159 int initial_cpb_removal_delay_offset;
160 double nrt_first_access_unit; /* nominal removal time */
161 double previous_cpb_final_arrival_time;
165 static int parse_zones( x264_t *h );
166 static int init_pass2(x264_t *);
167 static float rate_estimate_qscale( x264_t *h );
168 static int update_vbv( x264_t *h, int bits );
169 static void update_vbv_plan( x264_t *h, int overhead );
170 static double predict_size( predictor_t *p, double q, double var );
171 static void update_predictor( predictor_t *p, double q, double var, double bits );
173 #define CMP_OPT_FIRST_PASS( opt, param_val )\
175 if( ( p = strstr( opts, opt "=" ) ) && sscanf( p, opt "=%d" , &i ) && param_val != i )\
177 x264_log( h, X264_LOG_ERROR, "different " opt " setting than first pass (%d vs %d)\n", param_val, i );\
183 * qp = h.264's quantizer
184 * qscale = linearized quantizer = Lagrange multiplier
186 static inline double qp2qscale( double qp )
188 return 0.85 * pow( 2.0, ( qp - 12.0 ) / 6.0 );
190 static inline double qscale2qp( double qscale )
192 return 12.0 + 6.0 * log2( qscale/0.85 );
195 /* Texture bitrate is not quite inversely proportional to qscale,
196 * probably due the the changing number of SKIP blocks.
197 * MV bits level off at about qp<=12, because the lambda used
198 * for motion estimation is constant there. */
199 static inline double qscale2bits( ratecontrol_entry_t *rce, double qscale )
203 return (rce->tex_bits + .1) * pow( rce->qscale / qscale, 1.1 )
204 + rce->mv_bits * pow( X264_MAX(rce->qscale, 1) / X264_MAX(qscale, 1), 0.5 )
208 static ALWAYS_INLINE uint32_t ac_energy_plane( x264_t *h, int mb_x, int mb_y, x264_frame_t *frame, int i )
211 int shift = i ? 6 : 8;
212 int stride = frame->i_stride[i];
213 int offset = h->mb.b_interlaced
214 ? w * (mb_x + (mb_y&~1) * stride) + (mb_y&1) * stride
215 : w * (mb_x + mb_y * stride);
216 int pix = i ? PIXEL_8x8 : PIXEL_16x16;
217 stride <<= h->mb.b_interlaced;
218 uint64_t res = h->pixf.var[pix]( frame->plane[i] + offset, stride );
219 uint32_t sum = (uint32_t)res;
220 uint32_t sqr = res >> 32;
221 return sqr - (sum * sum >> shift);
224 // Find the total AC energy of the block in all planes.
225 static NOINLINE uint32_t ac_energy_mb( x264_t *h, int mb_x, int mb_y, x264_frame_t *frame )
227 /* This function contains annoying hacks because GCC has a habit of reordering emms
228 * and putting it after floating point ops. As a result, we put the emms at the end of the
229 * function and make sure that its always called before the float math. Noinline makes
230 * sure no reordering goes on. */
231 uint32_t var = ac_energy_plane( h, mb_x, mb_y, frame, 0 );
232 var += ac_energy_plane( h, mb_x, mb_y, frame, 1 );
233 var += ac_energy_plane( h, mb_x, mb_y, frame, 2 );
238 void x264_adaptive_quant_frame( x264_t *h, x264_frame_t *frame )
240 /* constants chosen to result in approximately the same overall bitrate as without AQ.
241 * FIXME: while they're written in 5 significant digits, they're only tuned to 2. */
244 /* Need to init it anyways for MB tree. */
245 if( h->param.rc.f_aq_strength == 0 )
247 memset( frame->f_qp_offset, 0, h->mb.i_mb_count * sizeof(float) );
248 memset( frame->f_qp_offset_aq, 0, h->mb.i_mb_count * sizeof(float) );
249 if( h->frames.b_have_lowres )
250 for( int mb_xy = 0; mb_xy < h->mb.i_mb_count; mb_xy++ )
251 frame->i_inv_qscale_factor[mb_xy] = 256;
255 if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE )
257 float avg_adj_pow2 = 0.f;
258 for( int mb_y = 0; mb_y < h->sps->i_mb_height; mb_y++ )
259 for( int mb_x = 0; mb_x < h->sps->i_mb_width; mb_x++ )
261 uint32_t energy = ac_energy_mb( h, mb_x, mb_y, frame );
262 float qp_adj = powf( energy + 1, 0.125f );
263 frame->f_qp_offset[mb_x + mb_y*h->mb.i_mb_stride] = qp_adj;
265 avg_adj_pow2 += qp_adj * qp_adj;
267 avg_adj /= h->mb.i_mb_count;
268 avg_adj_pow2 /= h->mb.i_mb_count;
269 strength = h->param.rc.f_aq_strength * avg_adj;
270 avg_adj = avg_adj - 0.5f * (avg_adj_pow2 - 14.f) / avg_adj;
273 strength = h->param.rc.f_aq_strength * 1.0397f;
275 for( int mb_y = 0; mb_y < h->sps->i_mb_height; mb_y++ )
276 for( int mb_x = 0; mb_x < h->sps->i_mb_width; mb_x++ )
279 if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE )
281 qp_adj = frame->f_qp_offset[mb_x + mb_y*h->mb.i_mb_stride];
282 qp_adj = strength * (qp_adj - avg_adj);
286 uint32_t energy = ac_energy_mb( h, mb_x, mb_y, frame );
287 qp_adj = strength * (x264_log2( X264_MAX(energy, 1) ) - 14.427f);
289 frame->f_qp_offset[mb_x + mb_y*h->mb.i_mb_stride] =
290 frame->f_qp_offset_aq[mb_x + mb_y*h->mb.i_mb_stride] = qp_adj;
291 if( h->frames.b_have_lowres )
292 frame->i_inv_qscale_factor[mb_x + mb_y*h->mb.i_mb_stride] = x264_exp2fix8(qp_adj);
297 /*****************************************************************************
298 * x264_adaptive_quant:
299 * adjust macroblock QP based on variance (AC energy) of the MB.
300 * high variance = higher QP
301 * low variance = lower QP
302 * This generally increases SSIM and lowers PSNR.
303 *****************************************************************************/
304 void x264_adaptive_quant( x264_t *h )
307 /* MB-tree currently doesn't adjust quantizers in unreferenced frames. */
308 float qp_offset = h->fdec->b_kept_as_ref ? h->fenc->f_qp_offset[h->mb.i_mb_xy] : h->fenc->f_qp_offset_aq[h->mb.i_mb_xy];
309 h->mb.i_qp = x264_clip3( h->rc->f_qpm + qp_offset + .5, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
312 int x264_macroblock_tree_read( x264_t *h, x264_frame_t *frame )
314 x264_ratecontrol_t *rc = h->rc;
315 uint8_t i_type_actual = rc->entry[frame->i_frame].pict_type;
317 if( rc->entry[frame->i_frame].kept_as_ref )
320 if( rc->qpbuf_pos < 0 )
326 if( !fread( &i_type, 1, 1, rc->p_mbtree_stat_file_in ) )
328 if( fread( rc->qp_buffer[rc->qpbuf_pos], sizeof(uint16_t), h->mb.i_mb_count, rc->p_mbtree_stat_file_in ) != h->mb.i_mb_count )
331 if( i_type != i_type_actual && rc->qpbuf_pos == 1 )
333 x264_log(h, X264_LOG_ERROR, "MB-tree frametype %d doesn't match actual frametype %d.\n", i_type, i_type_actual);
336 } while( i_type != i_type_actual );
339 for( int i = 0; i < h->mb.i_mb_count; i++ )
341 frame->f_qp_offset[i] = ((float)(int16_t)endian_fix16( rc->qp_buffer[rc->qpbuf_pos][i] )) * (1/256.0);
342 if( h->frames.b_have_lowres )
343 frame->i_inv_qscale_factor[i] = x264_exp2fix8(frame->f_qp_offset[i]);
348 x264_adaptive_quant_frame( h, frame );
351 x264_log(h, X264_LOG_ERROR, "Incomplete MB-tree stats file.\n");
355 int x264_reference_build_list_optimal( x264_t *h )
357 ratecontrol_entry_t *rce = h->rc->rce;
358 x264_frame_t *frames[16];
359 x264_weight_t weights[16][3];
362 if( rce->refs != h->i_ref0 )
365 memcpy( frames, h->fref0, sizeof(frames) );
366 memcpy( refcount, rce->refcount, sizeof(refcount) );
367 memcpy( weights, h->fenc->weight, sizeof(weights) );
368 memset( &h->fenc->weight[1][0], 0, sizeof(x264_weight_t[15][3]) );
370 /* For now don't reorder ref 0; it seems to lower quality
371 in most cases due to skips. */
372 for( int ref = 1; ref < h->i_ref0; ref++ )
377 for( int i = 1; i < h->i_ref0; i++ )
378 if( !frames[i]->b_duplicate || frames[i]->i_frame != h->fref0[ref-1]->i_frame )
379 /* Favor lower POC as a tiebreaker. */
380 COPY2_IF_GT( max, refcount[i], bestref, i );
382 /* FIXME: If there are duplicates from frames other than ref0 then it is possible
383 * that the optimal ordering doesnt place every duplicate. */
385 refcount[bestref] = -1;
386 h->fref0[ref] = frames[bestref];
387 memcpy( h->fenc->weight[ref], weights[bestref], sizeof(weights[bestref]) );
393 static char *x264_strcat_filename( char *input, char *suffix )
395 char *output = x264_malloc( strlen( input ) + strlen( suffix ) + 1 );
398 strcpy( output, input );
399 strcat( output, suffix );
403 void x264_ratecontrol_init_reconfigurable( x264_t *h, int b_init )
405 x264_ratecontrol_t *rc = h->rc;
406 if( !b_init && rc->b_2pass )
409 if( h->param.rc.i_rc_method == X264_RC_CRF )
411 /* Arbitrary rescaling to make CRF somewhat similar to QP.
412 * Try to compensate for MB-tree's effects as well. */
413 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
414 double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*13.5 : 0;
415 rc->rate_factor_constant = pow( base_cplx, 1 - rc->qcompress )
416 / qp2qscale( h->param.rc.f_rf_constant + mbtree_offset );
419 if( h->param.rc.i_vbv_max_bitrate > 0 && h->param.rc.i_vbv_buffer_size > 0 )
421 if( h->param.rc.i_vbv_buffer_size < (int)(h->param.rc.i_vbv_max_bitrate / rc->fps) )
423 h->param.rc.i_vbv_buffer_size = h->param.rc.i_vbv_max_bitrate / rc->fps;
424 x264_log( h, X264_LOG_WARNING, "VBV buffer size cannot be smaller than one frame, using %d kbit\n",
425 h->param.rc.i_vbv_buffer_size );
428 /* We don't support changing the ABR bitrate right now,
429 so if the stream starts as CBR, keep it CBR. */
430 if( rc->b_vbv_min_rate )
431 h->param.rc.i_vbv_max_bitrate = h->param.rc.i_bitrate;
433 int vbv_buffer_size = h->param.rc.i_vbv_buffer_size * 1000;
434 int vbv_max_bitrate = h->param.rc.i_vbv_max_bitrate * 1000;
437 if( h->param.i_nal_hrd && b_init )
439 h->sps->vui.hrd.i_cpb_cnt = 1;
440 h->sps->vui.hrd.b_cbr_hrd = h->param.i_nal_hrd == X264_NAL_HRD_CBR;
441 h->sps->vui.hrd.i_time_offset_length = 0;
446 int bitrate = 1000*h->param.rc.i_vbv_max_bitrate;
447 int bufsize = 1000*h->param.rc.i_vbv_buffer_size;
449 // normalize HRD size and rate to the value / scale notation
450 h->sps->vui.hrd.i_bit_rate_scale = x264_clip3( x264_ctz( bitrate ) - BR_SHIFT, 0, 15 );
451 h->sps->vui.hrd.i_bit_rate_value = bitrate >> ( h->sps->vui.hrd.i_bit_rate_scale + BR_SHIFT );
452 h->sps->vui.hrd.i_bit_rate_unscaled = h->sps->vui.hrd.i_bit_rate_value << ( h->sps->vui.hrd.i_bit_rate_scale + BR_SHIFT );
453 h->sps->vui.hrd.i_cpb_size_scale = x264_clip3( x264_ctz( bufsize ) - CPB_SHIFT, 0, 15 );
454 h->sps->vui.hrd.i_cpb_size_value = bufsize >> ( h->sps->vui.hrd.i_cpb_size_scale + CPB_SHIFT );
455 h->sps->vui.hrd.i_cpb_size_unscaled = h->sps->vui.hrd.i_cpb_size_value << ( h->sps->vui.hrd.i_cpb_size_scale + CPB_SHIFT );
461 #define MAX_DURATION 0.5
463 int max_cpb_output_delay = h->param.i_keyint_max * MAX_DURATION * h->sps->vui.i_time_scale / h->sps->vui.i_num_units_in_tick;
464 int max_dpb_output_delay = h->sps->vui.i_max_dec_frame_buffering * MAX_DURATION * h->sps->vui.i_time_scale / h->sps->vui.i_num_units_in_tick;
465 int max_delay = (int)(90000.0 * (double)h->sps->vui.hrd.i_cpb_size_unscaled / h->sps->vui.hrd.i_bit_rate_unscaled + 0.5);
467 h->sps->vui.hrd.i_initial_cpb_removal_delay_length = 2 + x264_clip3( 32 - x264_clz( max_delay ), 4, 22 );
468 h->sps->vui.hrd.i_cpb_removal_delay_length = x264_clip3( 32 - x264_clz( max_cpb_output_delay ), 4, 32 );
469 h->sps->vui.hrd.i_dpb_output_delay_length = x264_clip3( 32 - x264_clz( max_dpb_output_delay ), 4, 32 );
473 vbv_buffer_size = X264_MIN( vbv_buffer_size, h->sps->vui.hrd.i_cpb_size_unscaled );
474 vbv_max_bitrate = X264_MIN( vbv_max_bitrate, h->sps->vui.hrd.i_bit_rate_unscaled );
476 else if( h->param.i_nal_hrd && !b_init )
478 x264_log( h, X264_LOG_WARNING, "VBV parameters cannot be changed when NAL HRD is in use\n" );
482 rc->buffer_rate = vbv_max_bitrate / rc->fps;
483 rc->vbv_max_rate = vbv_max_bitrate;
484 rc->buffer_size = vbv_buffer_size;
485 rc->single_frame_vbv = rc->buffer_rate * 1.1 > rc->buffer_size;
486 rc->cbr_decay = 1.0 - rc->buffer_rate / rc->buffer_size
487 * 0.5 * X264_MAX(0, 1.5 - rc->buffer_rate * rc->fps / rc->bitrate);
488 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.f_rf_constant_max )
490 rc->rate_factor_max_increment = h->param.rc.f_rf_constant_max - h->param.rc.f_rf_constant;
491 if( rc->rate_factor_max_increment <= 0 )
493 x264_log( h, X264_LOG_WARNING, "CRF max must be greater than CRF\n" );
494 rc->rate_factor_max_increment = 0;
499 if( h->param.rc.f_vbv_buffer_init > 1. )
500 h->param.rc.f_vbv_buffer_init = x264_clip3f( h->param.rc.f_vbv_buffer_init / h->param.rc.i_vbv_buffer_size, 0, 1 );
501 h->param.rc.f_vbv_buffer_init = x264_clip3f( X264_MAX( h->param.rc.f_vbv_buffer_init, rc->buffer_rate / rc->buffer_size ), 0, 1);
502 rc->buffer_fill_final = rc->buffer_size * h->param.rc.f_vbv_buffer_init;
504 rc->b_vbv_min_rate = !rc->b_2pass
505 && h->param.rc.i_rc_method == X264_RC_ABR
506 && h->param.rc.i_vbv_max_bitrate <= h->param.rc.i_bitrate;
511 int x264_ratecontrol_new( x264_t *h )
513 x264_ratecontrol_t *rc;
517 CHECKED_MALLOCZERO( h->rc, h->param.i_threads * sizeof(x264_ratecontrol_t) );
520 rc->b_abr = h->param.rc.i_rc_method != X264_RC_CQP && !h->param.rc.b_stat_read;
521 rc->b_2pass = h->param.rc.i_rc_method == X264_RC_ABR && h->param.rc.b_stat_read;
523 /* FIXME: use integers */
524 if( h->param.i_fps_num > 0 && h->param.i_fps_den > 0 )
525 rc->fps = (float) h->param.i_fps_num / h->param.i_fps_den;
529 if( h->param.rc.b_mb_tree )
531 h->param.rc.f_pb_factor = 1;
535 rc->qcompress = h->param.rc.f_qcompress;
537 rc->bitrate = h->param.rc.i_bitrate * 1000.;
538 rc->rate_tolerance = h->param.rc.f_rate_tolerance;
539 rc->nmb = h->mb.i_mb_count;
540 rc->last_non_b_pict_type = -1;
543 if( h->param.rc.i_rc_method == X264_RC_CRF && h->param.rc.b_stat_read )
545 x264_log(h, X264_LOG_ERROR, "constant rate-factor is incompatible with 2pass.\n");
549 x264_ratecontrol_init_reconfigurable( h, 1 );
551 if( rc->rate_tolerance < 0.01 )
553 x264_log(h, X264_LOG_WARNING, "bitrate tolerance too small, using .01\n");
554 rc->rate_tolerance = 0.01;
557 h->mb.b_variable_qp = rc->b_vbv || h->param.rc.i_aq_mode;
561 /* FIXME ABR_INIT_QP is actually used only in CRF */
562 #define ABR_INIT_QP ( h->param.rc.i_rc_method == X264_RC_CRF ? h->param.rc.f_rf_constant : 24 )
563 rc->accum_p_norm = .01;
564 rc->accum_p_qp = ABR_INIT_QP * rc->accum_p_norm;
565 /* estimated ratio that produces a reasonable QP for the first I-frame */
566 rc->cplxr_sum = .01 * pow( 7.0e5, rc->qcompress ) * pow( h->mb.i_mb_count, 0.5 );
567 rc->wanted_bits_window = 1.0 * rc->bitrate / rc->fps;
568 rc->last_non_b_pict_type = SLICE_TYPE_I;
571 rc->ip_offset = 6.0 * log2f( h->param.rc.f_ip_factor );
572 rc->pb_offset = 6.0 * log2f( h->param.rc.f_pb_factor );
573 rc->qp_constant[SLICE_TYPE_P] = h->param.rc.i_qp_constant;
574 rc->qp_constant[SLICE_TYPE_I] = x264_clip3( h->param.rc.i_qp_constant - rc->ip_offset + 0.5, 0, 51 );
575 rc->qp_constant[SLICE_TYPE_B] = x264_clip3( h->param.rc.i_qp_constant + rc->pb_offset + 0.5, 0, 51 );
576 h->mb.ip_offset = rc->ip_offset + 0.5;
578 rc->lstep = pow( 2, h->param.rc.i_qp_step / 6.0 );
579 rc->last_qscale = qp2qscale( 26 );
580 int num_preds = h->param.b_sliced_threads * h->param.i_threads + 1;
581 CHECKED_MALLOC( rc->pred, 5 * sizeof(predictor_t) * num_preds );
582 CHECKED_MALLOC( rc->pred_b_from_p, sizeof(predictor_t) );
583 for( int i = 0; i < 5; i++ )
585 rc->last_qscale_for[i] = qp2qscale( ABR_INIT_QP );
586 rc->lmin[i] = qp2qscale( h->param.rc.i_qp_min );
587 rc->lmax[i] = qp2qscale( h->param.rc.i_qp_max );
588 for( int j = 0; j < num_preds; j++ )
590 rc->pred[i+j*5].coeff= 2.0;
591 rc->pred[i+j*5].count= 1.0;
592 rc->pred[i+j*5].decay= 0.5;
593 rc->pred[i+j*5].offset= 0.0;
595 for( int j = 0; j < 2; j++ )
597 rc->row_preds[i][j].coeff= .25;
598 rc->row_preds[i][j].count= 1.0;
599 rc->row_preds[i][j].decay= 0.5;
600 rc->row_preds[i][j].offset= 0.0;
603 *rc->pred_b_from_p = rc->pred[0];
605 if( parse_zones( h ) < 0 )
607 x264_log( h, X264_LOG_ERROR, "failed to parse zones\n" );
611 /* Load stat file and init 2pass algo */
612 if( h->param.rc.b_stat_read )
614 char *p, *stats_in, *stats_buf;
616 /* read 1st pass stats */
617 assert( h->param.rc.psz_stat_in );
618 stats_buf = stats_in = x264_slurp_file( h->param.rc.psz_stat_in );
621 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
624 if( h->param.rc.b_mb_tree )
626 char *mbtree_stats_in = x264_strcat_filename( h->param.rc.psz_stat_in, ".mbtree" );
627 if( !mbtree_stats_in )
629 rc->p_mbtree_stat_file_in = fopen( mbtree_stats_in, "rb" );
630 x264_free( mbtree_stats_in );
631 if( !rc->p_mbtree_stat_file_in )
633 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n");
638 /* check whether 1st pass options were compatible with current options */
639 if( !strncmp( stats_buf, "#options:", 9 ) )
643 char *opts = stats_buf;
644 stats_in = strchr( stats_buf, '\n' );
649 if( sscanf( opts, "#options: %dx%d", &i, &j ) != 2 )
651 x264_log( h, X264_LOG_ERROR, "resolution specified in stats file not valid\n" );
654 else if( h->param.rc.b_mb_tree && (i != h->param.i_width || j != h->param.i_height) )
656 x264_log( h, X264_LOG_ERROR, "MB-tree doesn't support different resolution than 1st pass (%dx%d vs %dx%d)\n",
657 h->param.i_width, h->param.i_height, i, j );
661 if( ( p = strstr( opts, "timebase=" ) ) && sscanf( p, "timebase=%u/%u", &k, &l ) != 2 )
663 x264_log( h, X264_LOG_ERROR, "timebase specified in stats file not valid\n" );
666 if( k != h->param.i_timebase_num || l != h->param.i_timebase_den )
668 x264_log( h, X264_LOG_ERROR, "timebase mismatch with 1st pass (%u/%u vs %u/%u)\n",
669 h->param.i_timebase_num, h->param.i_timebase_den, k, l );
673 CMP_OPT_FIRST_PASS( "wpredp", X264_MAX( 0, h->param.analyse.i_weighted_pred ) );
674 CMP_OPT_FIRST_PASS( "bframes", h->param.i_bframe );
675 CMP_OPT_FIRST_PASS( "b_pyramid", h->param.i_bframe_pyramid );
676 CMP_OPT_FIRST_PASS( "intra_refresh", h->param.b_intra_refresh );
677 CMP_OPT_FIRST_PASS( "keyint", h->param.i_keyint_max );
679 if( strstr( opts, "qp=0" ) && h->param.rc.i_rc_method == X264_RC_ABR )
680 x264_log( h, X264_LOG_WARNING, "1st pass was lossless, bitrate prediction will be inaccurate\n" );
682 if( !strstr( opts, "direct=3" ) && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
684 x264_log( h, X264_LOG_WARNING, "direct=auto not used on the first pass\n" );
685 h->mb.b_direct_auto_write = 1;
688 if( ( p = strstr( opts, "b_adapt=" ) ) && sscanf( p, "b_adapt=%d", &i ) && i >= X264_B_ADAPT_NONE && i <= X264_B_ADAPT_TRELLIS )
689 h->param.i_bframe_adaptive = i;
690 else if( h->param.i_bframe )
692 x264_log( h, X264_LOG_ERROR, "b_adapt method specified in stats file not valid\n" );
696 if( h->param.rc.b_mb_tree && ( p = strstr( opts, "rc_lookahead=" ) ) && sscanf( p, "rc_lookahead=%d", &i ) )
697 h->param.rc.i_lookahead = i;
700 /* find number of pics */
703 for( num_entries = -1; p; num_entries++ )
704 p = strchr( p + 1, ';' );
707 x264_log(h, X264_LOG_ERROR, "empty stats file\n");
710 rc->num_entries = num_entries;
712 if( h->param.i_frame_total < rc->num_entries && h->param.i_frame_total > 0 )
714 x264_log( h, X264_LOG_WARNING, "2nd pass has fewer frames than 1st pass (%d vs %d)\n",
715 h->param.i_frame_total, rc->num_entries );
717 if( h->param.i_frame_total > rc->num_entries )
719 x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d vs %d)\n",
720 h->param.i_frame_total, rc->num_entries );
724 CHECKED_MALLOCZERO( rc->entry, rc->num_entries * sizeof(ratecontrol_entry_t) );
726 /* init all to skipped p frames */
727 for( int i = 0; i < rc->num_entries; i++ )
729 ratecontrol_entry_t *rce = &rc->entry[i];
730 rce->pict_type = SLICE_TYPE_P;
731 rce->qscale = rce->new_qscale = qp2qscale( 20 );
732 rce->misc_bits = rc->nmb + 10;
738 for( int i = 0; i < rc->num_entries; i++ )
740 ratecontrol_entry_t *rce;
748 next= strchr(p, ';');
750 *next++ = 0; //sscanf is unbelievably slow on long strings
751 e = sscanf( p, " in:%d ", &frame_number );
753 if( frame_number < 0 || frame_number >= rc->num_entries )
755 x264_log( h, X264_LOG_ERROR, "bad frame number (%d) at stats line %d\n", frame_number, i );
758 rce = &rc->entry[frame_number];
759 rce->direct_mode = 0;
761 e += sscanf( p, " in:%*d out:%*d type:%c dur:%d cpbdur:%d q:%f tex:%d mv:%d misc:%d imb:%d pmb:%d smb:%d d:%c",
762 &pict_type, &rce->i_duration, &rce->i_cpb_duration, &qp, &rce->tex_bits,
763 &rce->mv_bits, &rce->misc_bits, &rce->i_count, &rce->p_count,
764 &rce->s_count, &rce->direct_mode );
766 p = strstr( p, "ref:" );
770 for( ref = 0; ref < 16; ref++ )
772 if( sscanf( p, " %d", &rce->refcount[ref] ) != 1 )
774 p = strchr( p+1, ' ' );
781 rce->i_weight_denom = -1;
782 char *w = strchr( p, 'w' );
784 if( sscanf( w, "w:%hd,%hd,%hd", &rce->i_weight_denom, &rce->weight[0], &rce->weight[1] ) != 3 )
785 rce->i_weight_denom = -1;
787 if( pict_type != 'b' )
788 rce->kept_as_ref = 1;
792 rce->frame_type = X264_TYPE_IDR;
793 rce->pict_type = SLICE_TYPE_I;
796 rce->frame_type = X264_TYPE_I;
797 rce->pict_type = SLICE_TYPE_I;
800 rce->frame_type = X264_TYPE_P;
801 rce->pict_type = SLICE_TYPE_P;
804 rce->frame_type = X264_TYPE_BREF;
805 rce->pict_type = SLICE_TYPE_B;
808 rce->frame_type = X264_TYPE_B;
809 rce->pict_type = SLICE_TYPE_B;
811 default: e = -1; break;
816 x264_log( h, X264_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e );
819 rce->qscale = qp2qscale( qp );
823 x264_free( stats_buf );
825 if( h->param.rc.i_rc_method == X264_RC_ABR )
827 if( init_pass2( h ) < 0 )
829 } /* else we're using constant quant, so no need to run the bitrate allocation */
832 /* Open output file */
833 /* If input and output files are the same, output to a temp file
834 * and move it to the real name only when it's complete */
835 if( h->param.rc.b_stat_write )
838 rc->psz_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".temp" );
839 if( !rc->psz_stat_file_tmpname )
842 rc->p_stat_file_out = fopen( rc->psz_stat_file_tmpname, "wb" );
843 if( rc->p_stat_file_out == NULL )
845 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n");
849 p = x264_param2string( &h->param, 1 );
851 fprintf( rc->p_stat_file_out, "#options: %s\n", p );
853 if( h->param.rc.b_mb_tree && !h->param.rc.b_stat_read )
855 rc->psz_mbtree_stat_file_tmpname = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree.temp" );
856 rc->psz_mbtree_stat_file_name = x264_strcat_filename( h->param.rc.psz_stat_out, ".mbtree" );
857 if( !rc->psz_mbtree_stat_file_tmpname || !rc->psz_mbtree_stat_file_name )
860 rc->p_mbtree_stat_file_out = fopen( rc->psz_mbtree_stat_file_tmpname, "wb" );
861 if( rc->p_mbtree_stat_file_out == NULL )
863 x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open mbtree stats file\n");
869 if( h->param.rc.b_mb_tree && (h->param.rc.b_stat_read || h->param.rc.b_stat_write) )
871 CHECKED_MALLOC( rc->qp_buffer[0], h->mb.i_mb_count * sizeof(uint16_t) );
872 if( h->param.i_bframe_pyramid && h->param.rc.b_stat_read )
873 CHECKED_MALLOC( rc->qp_buffer[1], h->mb.i_mb_count * sizeof(uint16_t) );
877 for( int i = 0; i<h->param.i_threads; i++ )
879 h->thread[i]->rc = rc+i;
883 h->thread[i]->param = h->param;
884 h->thread[i]->mb.b_variable_qp = h->mb.b_variable_qp;
893 static int parse_zone( x264_t *h, x264_zone_t *z, char *p )
896 char *tok, UNUSED *saveptr=NULL;
898 z->f_bitrate_factor = 1;
899 if( 3 <= sscanf(p, "%u,%u,q=%u%n", &z->i_start, &z->i_end, &z->i_qp, &len) )
901 else if( 3 <= sscanf(p, "%u,%u,b=%f%n", &z->i_start, &z->i_end, &z->f_bitrate_factor, &len) )
903 else if( 2 <= sscanf(p, "%u,%u%n", &z->i_start, &z->i_end, &len) )
907 x264_log( h, X264_LOG_ERROR, "invalid zone: \"%s\"\n", p );
913 CHECKED_MALLOC( z->param, sizeof(x264_param_t) );
914 memcpy( z->param, &h->param, sizeof(x264_param_t) );
915 z->param->param_free = x264_free;
916 while( (tok = strtok_r( p, ",", &saveptr )) )
918 char *val = strchr( tok, '=' );
924 if( x264_param_parse( z->param, tok, val ) )
926 x264_log( h, X264_LOG_ERROR, "invalid zone param: %s = %s\n", tok, val );
936 static int parse_zones( x264_t *h )
938 x264_ratecontrol_t *rc = h->rc;
939 if( h->param.rc.psz_zones && !h->param.rc.i_zones )
942 CHECKED_MALLOC( psz_zones, strlen( h->param.rc.psz_zones )+1 );
943 strcpy( psz_zones, h->param.rc.psz_zones );
944 h->param.rc.i_zones = 1;
945 for( p = psz_zones; *p; p++ )
946 h->param.rc.i_zones += (*p == '/');
947 CHECKED_MALLOC( h->param.rc.zones, h->param.rc.i_zones * sizeof(x264_zone_t) );
949 for( int i = 0; i < h->param.rc.i_zones; i++ )
951 int i_tok = strcspn( p, "/" );
953 if( parse_zone( h, &h->param.rc.zones[i], p ) )
957 x264_free( psz_zones );
960 if( h->param.rc.i_zones > 0 )
962 for( int i = 0; i < h->param.rc.i_zones; i++ )
964 x264_zone_t z = h->param.rc.zones[i];
965 if( z.i_start < 0 || z.i_start > z.i_end )
967 x264_log( h, X264_LOG_ERROR, "invalid zone: start=%d end=%d\n",
968 z.i_start, z.i_end );
971 else if( !z.b_force_qp && z.f_bitrate_factor <= 0 )
973 x264_log( h, X264_LOG_ERROR, "invalid zone: bitrate_factor=%f\n",
974 z.f_bitrate_factor );
979 rc->i_zones = h->param.rc.i_zones + 1;
980 CHECKED_MALLOC( rc->zones, rc->i_zones * sizeof(x264_zone_t) );
981 memcpy( rc->zones+1, h->param.rc.zones, (rc->i_zones-1) * sizeof(x264_zone_t) );
983 // default zone to fall back to if none of the others match
984 rc->zones[0].i_start = 0;
985 rc->zones[0].i_end = INT_MAX;
986 rc->zones[0].b_force_qp = 0;
987 rc->zones[0].f_bitrate_factor = 1;
988 CHECKED_MALLOC( rc->zones[0].param, sizeof(x264_param_t) );
989 memcpy( rc->zones[0].param, &h->param, sizeof(x264_param_t) );
990 for( int i = 1; i < rc->i_zones; i++ )
992 if( !rc->zones[i].param )
993 rc->zones[i].param = rc->zones[0].param;
1002 static x264_zone_t *get_zone( x264_t *h, int frame_num )
1004 for( int i = h->rc->i_zones - 1; i >= 0; i-- )
1006 x264_zone_t *z = &h->rc->zones[i];
1007 if( frame_num >= z->i_start && frame_num <= z->i_end )
1013 void x264_ratecontrol_summary( x264_t *h )
1015 x264_ratecontrol_t *rc = h->rc;
1016 if( rc->b_abr && h->param.rc.i_rc_method == X264_RC_ABR && rc->cbr_decay > .9999 )
1018 double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80);
1019 double mbtree_offset = h->param.rc.b_mb_tree ? (1.0-h->param.rc.f_qcompress)*13.5 : 0;
1020 x264_log( h, X264_LOG_INFO, "final ratefactor: %.2f\n",
1021 qscale2qp( pow( base_cplx, 1 - rc->qcompress )
1022 * rc->cplxr_sum / rc->wanted_bits_window ) - mbtree_offset );
1026 void x264_ratecontrol_delete( x264_t *h )
1028 x264_ratecontrol_t *rc = h->rc;
1031 if( rc->p_stat_file_out )
1033 b_regular_file = x264_is_regular_file( rc->p_stat_file_out );
1034 fclose( rc->p_stat_file_out );
1035 if( h->i_frame >= rc->num_entries && b_regular_file )
1036 if( rename( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out ) != 0 )
1038 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
1039 rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out );
1041 x264_free( rc->psz_stat_file_tmpname );
1043 if( rc->p_mbtree_stat_file_out )
1045 b_regular_file = x264_is_regular_file( rc->p_mbtree_stat_file_out );
1046 fclose( rc->p_mbtree_stat_file_out );
1047 if( h->i_frame >= rc->num_entries && b_regular_file )
1048 if( rename( rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name ) != 0 )
1050 x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n",
1051 rc->psz_mbtree_stat_file_tmpname, rc->psz_mbtree_stat_file_name );
1053 x264_free( rc->psz_mbtree_stat_file_tmpname );
1054 x264_free( rc->psz_mbtree_stat_file_name );
1056 if( rc->p_mbtree_stat_file_in )
1057 fclose( rc->p_mbtree_stat_file_in );
1058 x264_free( rc->pred );
1059 x264_free( rc->pred_b_from_p );
1060 x264_free( rc->entry );
1061 x264_free( rc->qp_buffer[0] );
1062 x264_free( rc->qp_buffer[1] );
1065 x264_free( rc->zones[0].param );
1066 for( int i = 1; i < rc->i_zones; i++ )
1067 if( rc->zones[i].param != rc->zones[0].param && rc->zones[i].param->param_free )
1068 rc->zones[i].param->param_free( rc->zones[i].param );
1069 x264_free( rc->zones );
1074 static void accum_p_qp_update( x264_t *h, float qp )
1076 x264_ratecontrol_t *rc = h->rc;
1077 rc->accum_p_qp *= .95;
1078 rc->accum_p_norm *= .95;
1079 rc->accum_p_norm += 1;
1080 if( h->sh.i_type == SLICE_TYPE_I )
1081 rc->accum_p_qp += qp + rc->ip_offset;
1083 rc->accum_p_qp += qp;
1086 /* Before encoding a frame, choose a QP for it */
1087 void x264_ratecontrol_start( x264_t *h, int i_force_qp, int overhead )
1089 x264_ratecontrol_t *rc = h->rc;
1090 ratecontrol_entry_t *rce = NULL;
1091 x264_zone_t *zone = get_zone( h, h->fenc->i_frame );
1096 if( zone && (!rc->prev_zone || zone->param != rc->prev_zone->param) )
1097 x264_encoder_reconfig( h, zone->param );
1098 rc->prev_zone = zone;
1100 rc->qp_force = i_force_qp;
1102 if( h->param.rc.b_stat_read )
1104 int frame = h->fenc->i_frame;
1105 assert( frame >= 0 && frame < rc->num_entries );
1106 rce = h->rc->rce = &h->rc->entry[frame];
1108 if( h->sh.i_type == SLICE_TYPE_B
1109 && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO )
1111 h->sh.b_direct_spatial_mv_pred = ( rce->direct_mode == 's' );
1112 h->mb.b_direct_auto_read = ( rce->direct_mode == 's' || rce->direct_mode == 't' );
1118 memset( h->fdec->i_row_bits, 0, h->sps->i_mb_height * sizeof(int) );
1119 rc->row_pred = &rc->row_preds[h->sh.i_type];
1120 rc->buffer_rate = h->fenc->i_cpb_duration * rc->vbv_max_rate * h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1121 update_vbv_plan( h, overhead );
1123 const x264_level_t *l = x264_levels;
1124 while( l->level_idc != 0 && l->level_idc != h->param.i_level_idc )
1127 int mincr = l->mincr;
1129 /* Blu-ray requires this */
1130 if( l->level_idc == 41 && h->param.i_nal_hrd )
1133 /* The spec has a bizarre special case for the first frame. */
1134 if( h->i_frame == 0 )
1136 //384 * ( Max( PicSizeInMbs, fR * MaxMBPS ) + MaxMBPS * ( tr( 0 ) - tr,n( 0 ) ) ) / MinCR
1137 double fr = 1. / 172;
1138 int pic_size_in_mbs = h->sps->i_mb_width * h->sps->i_mb_height;
1139 rc->frame_size_maximum = 384 * 8 * X264_MAX( pic_size_in_mbs, fr*l->mbps ) / mincr;
1143 //384 * MaxMBPS * ( tr( n ) - tr( n - 1 ) ) / MinCR
1144 rc->frame_size_maximum = 384 * 8 * ((double)h->fenc->i_cpb_duration * h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale) * l->mbps / mincr;
1148 if( h->sh.i_type != SLICE_TYPE_B )
1149 rc->bframes = h->fenc->i_bframes;
1155 else if( rc->b_abr )
1157 q = qscale2qp( rate_estimate_qscale( h ) );
1159 else if( rc->b_2pass )
1161 rce->new_qscale = rate_estimate_qscale( h );
1162 q = qscale2qp( rce->new_qscale );
1166 if( h->sh.i_type == SLICE_TYPE_B && h->fdec->b_kept_as_ref )
1167 q = ( rc->qp_constant[ SLICE_TYPE_B ] + rc->qp_constant[ SLICE_TYPE_P ] ) / 2;
1169 q = rc->qp_constant[ h->sh.i_type ];
1173 if( zone->b_force_qp )
1174 q += zone->i_qp - rc->qp_constant[SLICE_TYPE_P];
1176 q -= 6*log2f( zone->f_bitrate_factor );
1180 q = x264_clip3f( q, h->param.rc.i_qp_min, h->param.rc.i_qp_max );
1185 rc->qp = x264_clip3( (int)(q + 0.5), 0, 51 );
1186 h->fdec->f_qp_avg_rc =
1187 h->fdec->f_qp_avg_aq =
1190 rce->new_qp = rc->qp;
1192 accum_p_qp_update( h, rc->f_qpm );
1194 if( h->sh.i_type != SLICE_TYPE_B )
1195 rc->last_non_b_pict_type = h->sh.i_type;
1198 static double predict_row_size( x264_t *h, int y, int qp )
1200 /* average between two predictors:
1201 * absolute SATD, and scaled bit cost of the colocated row in the previous frame */
1202 x264_ratecontrol_t *rc = h->rc;
1203 double pred_s = predict_size( rc->row_pred[0], qp2qscale( qp ), h->fdec->i_row_satd[y] );
1205 if( h->sh.i_type == SLICE_TYPE_I || qp >= h->fref0[0]->i_row_qp[y] )
1207 if( h->sh.i_type == SLICE_TYPE_P
1208 && h->fref0[0]->i_type == h->fdec->i_type
1209 && h->fref0[0]->i_row_satd[y] > 0
1210 && (abs(h->fref0[0]->i_row_satd[y] - h->fdec->i_row_satd[y]) < h->fdec->i_row_satd[y]/2))
1212 pred_t = h->fref0[0]->i_row_bits[y] * h->fdec->i_row_satd[y] / h->fref0[0]->i_row_satd[y]
1213 * qp2qscale( h->fref0[0]->i_row_qp[y] ) / qp2qscale( qp );
1217 return (pred_s + pred_t) / 2;
1219 /* Our QP is lower than the reference! */
1222 double pred_intra = predict_size( rc->row_pred[1], qp2qscale( qp ), h->fdec->i_row_satds[0][0][y] );
1223 /* Sum: better to overestimate than underestimate by using only one of the two predictors. */
1224 return pred_intra + pred_s;
1228 static double row_bits_so_far( x264_t *h, int y )
1231 for( int i = h->i_threadslice_start; i <= y; i++ )
1232 bits += h->fdec->i_row_bits[i];
1236 static double predict_row_size_sum( x264_t *h, int y, int qp )
1238 double bits = row_bits_so_far(h, y);
1239 for( int i = y+1; i < h->i_threadslice_end; i++ )
1240 bits += predict_row_size( h, i, qp );
1245 void x264_ratecontrol_mb( x264_t *h, int bits )
1247 x264_ratecontrol_t *rc = h->rc;
1248 const int y = h->mb.i_mb_y;
1252 h->fdec->i_row_bits[y] += bits;
1253 rc->qpa_rc += rc->f_qpm;
1254 rc->qpa_aq += h->mb.i_qp;
1256 if( h->mb.i_mb_x != h->sps->i_mb_width - 1 || !rc->b_vbv )
1259 h->fdec->i_row_qp[y] = rc->qpm;
1261 update_predictor( rc->row_pred[0], qp2qscale( rc->qpm ), h->fdec->i_row_satd[y], h->fdec->i_row_bits[y] );
1262 if( h->sh.i_type == SLICE_TYPE_P && rc->qpm < h->fref0[0]->i_row_qp[y] )
1263 update_predictor( rc->row_pred[1], qp2qscale( rc->qpm ), h->fdec->i_row_satds[0][0][y], h->fdec->i_row_bits[y] );
1265 /* tweak quality based on difference from predicted size */
1266 if( y < h->i_threadslice_end-1 )
1268 int prev_row_qp = h->fdec->i_row_qp[y];
1269 int i_qp_min = X264_MAX( prev_row_qp - h->param.rc.i_qp_step, h->param.rc.i_qp_min );
1270 int i_qp_absolute_max = h->param.rc.i_qp_max;
1271 if( rc->rate_factor_max_increment )
1272 i_qp_absolute_max = X264_MIN( i_qp_absolute_max, rc->qp_novbv + rc->rate_factor_max_increment );
1273 int i_qp_max = X264_MIN( prev_row_qp + h->param.rc.i_qp_step, i_qp_absolute_max );
1275 /* B-frames shouldn't use lower QP than their reference frames. */
1276 if( h->sh.i_type == SLICE_TYPE_B )
1278 i_qp_min = X264_MAX( i_qp_min, X264_MAX( h->fref0[0]->i_row_qp[y+1], h->fref1[0]->i_row_qp[y+1] ) );
1279 rc->qpm = X264_MAX( rc->qpm, i_qp_min );
1282 float buffer_left_planned = rc->buffer_fill - rc->frame_size_planned;
1283 float slice_size_planned = h->param.b_sliced_threads ? rc->slice_size_planned : rc->frame_size_planned;
1284 float size_of_other_slices = 0;
1285 if( h->param.b_sliced_threads )
1287 for( int i = 0; i < h->param.i_threads; i++ )
1288 if( h != h->thread[i] )
1289 size_of_other_slices += h->thread[i]->rc->frame_size_estimated;
1292 rc->max_frame_error = X264_MAX( 0.05, 1.0 / (h->sps->i_mb_width) );
1294 /* More threads means we have to be more cautious in letting ratecontrol use up extra bits. */
1295 float rc_tol = buffer_left_planned / h->param.i_threads * rc->rate_tolerance;
1296 int b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1298 /* Don't modify the row QPs until a sufficent amount of the bits of the frame have been processed, in case a flat */
1299 /* area at the top of the frame was measured inaccurately. */
1300 if( row_bits_so_far( h, y ) < 0.05 * slice_size_planned )
1303 if( h->sh.i_type != SLICE_TYPE_I )
1306 if( !rc->b_vbv_min_rate )
1307 i_qp_min = X264_MAX( i_qp_min, h->sh.i_qp );
1309 while( rc->qpm < i_qp_max
1310 && ((b1 > rc->frame_size_planned + rc_tol) ||
1311 (rc->buffer_fill - b1 < buffer_left_planned * 0.5) ||
1312 (b1 > rc->frame_size_planned && rc->qpm < rc->qp_novbv)) )
1315 b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1318 while( rc->qpm > i_qp_min
1319 && (rc->qpm > h->fdec->i_row_qp[0] || rc->single_frame_vbv)
1320 && ((b1 < rc->frame_size_planned * 0.8 && rc->qpm <= prev_row_qp)
1321 || b1 < (rc->buffer_fill - rc->buffer_size + rc->buffer_rate) * 1.1) )
1324 b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1327 /* avoid VBV underflow or MinCR violation */
1328 while( (rc->qpm < i_qp_absolute_max)
1329 && ((rc->buffer_fill - b1 < rc->buffer_rate * rc->max_frame_error) ||
1330 (rc->frame_size_maximum - b1 < rc->frame_size_maximum * rc->max_frame_error)))
1333 b1 = predict_row_size_sum( h, y, rc->qpm ) + size_of_other_slices;
1336 h->rc->frame_size_estimated = predict_row_size_sum( h, y, rc->qpm );
1339 /* loses the fractional part of the frame-wise qp */
1340 rc->f_qpm = rc->qpm;
1343 int x264_ratecontrol_qp( x264_t *h )
1348 /* In 2pass, force the same frame types as in the 1st pass */
1349 int x264_ratecontrol_slice_type( x264_t *h, int frame_num )
1351 x264_ratecontrol_t *rc = h->rc;
1352 if( h->param.rc.b_stat_read )
1354 if( frame_num >= rc->num_entries )
1356 /* We could try to initialize everything required for ABR and
1357 * adaptive B-frames, but that would be complicated.
1358 * So just calculate the average QP used so far. */
1359 h->param.rc.i_qp_constant = (h->stat.i_frame_count[SLICE_TYPE_P] == 0) ? 24
1360 : 1 + h->stat.f_frame_qp[SLICE_TYPE_P] / h->stat.i_frame_count[SLICE_TYPE_P];
1361 rc->qp_constant[SLICE_TYPE_P] = x264_clip3( h->param.rc.i_qp_constant, 0, 51 );
1362 rc->qp_constant[SLICE_TYPE_I] = x264_clip3( (int)( qscale2qp( qp2qscale( h->param.rc.i_qp_constant ) / fabs( h->param.rc.f_ip_factor )) + 0.5 ), 0, 51 );
1363 rc->qp_constant[SLICE_TYPE_B] = x264_clip3( (int)( qscale2qp( qp2qscale( h->param.rc.i_qp_constant ) * fabs( h->param.rc.f_pb_factor )) + 0.5 ), 0, 51 );
1365 x264_log(h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d)\n", rc->num_entries);
1366 x264_log(h, X264_LOG_ERROR, "continuing anyway, at constant QP=%d\n", h->param.rc.i_qp_constant);
1367 if( h->param.i_bframe_adaptive )
1368 x264_log(h, X264_LOG_ERROR, "disabling adaptive B-frames\n");
1370 for( int i = 0; i < h->param.i_threads; i++ )
1372 h->thread[i]->rc->b_abr = 0;
1373 h->thread[i]->rc->b_2pass = 0;
1374 h->thread[i]->param.rc.i_rc_method = X264_RC_CQP;
1375 h->thread[i]->param.rc.b_stat_read = 0;
1376 h->thread[i]->param.i_bframe_adaptive = 0;
1377 h->thread[i]->param.i_scenecut_threshold = 0;
1378 h->thread[i]->param.rc.b_mb_tree = 0;
1379 if( h->thread[i]->param.i_bframe > 1 )
1380 h->thread[i]->param.i_bframe = 1;
1382 return X264_TYPE_AUTO;
1384 return rc->entry[frame_num].frame_type;
1387 return X264_TYPE_AUTO;
1390 void x264_ratecontrol_set_weights( x264_t *h, x264_frame_t *frm )
1392 ratecontrol_entry_t *rce = &h->rc->entry[frm->i_frame];
1393 if( h->param.analyse.i_weighted_pred <= 0 )
1395 if( rce->i_weight_denom >= 0 )
1396 SET_WEIGHT( frm->weight[0][0], 1, rce->weight[0], rce->i_weight_denom, rce->weight[1] );
1399 /* After encoding one frame, save stats and update ratecontrol state */
1400 int x264_ratecontrol_end( x264_t *h, int bits, int *filler )
1402 x264_ratecontrol_t *rc = h->rc;
1403 const int *mbs = h->stat.frame.i_mb_count;
1407 h->stat.frame.i_mb_count_skip = mbs[P_SKIP] + mbs[B_SKIP];
1408 h->stat.frame.i_mb_count_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
1409 h->stat.frame.i_mb_count_p = mbs[P_L0] + mbs[P_8x8];
1410 for( int i = B_DIRECT; i < B_8x8; i++ )
1411 h->stat.frame.i_mb_count_p += mbs[i];
1413 h->fdec->f_qp_avg_rc = rc->qpa_rc /= h->mb.i_mb_count;
1414 h->fdec->f_qp_avg_aq = rc->qpa_aq /= h->mb.i_mb_count;
1416 if( h->param.rc.b_stat_write )
1418 char c_type = h->sh.i_type==SLICE_TYPE_I ? (h->fenc->i_poc==0 ? 'I' : 'i')
1419 : h->sh.i_type==SLICE_TYPE_P ? 'P'
1420 : h->fenc->b_kept_as_ref ? 'B' : 'b';
1421 int dir_frame = h->stat.frame.i_direct_score[1] - h->stat.frame.i_direct_score[0];
1422 int dir_avg = h->stat.i_direct_score[1] - h->stat.i_direct_score[0];
1423 char c_direct = h->mb.b_direct_auto_write ?
1424 ( dir_frame>0 ? 's' : dir_frame<0 ? 't' :
1425 dir_avg>0 ? 's' : dir_avg<0 ? 't' : '-' )
1427 if( fprintf( rc->p_stat_file_out,
1428 "in:%d out:%d type:%c dur:%d cpbdur:%d q:%.2f tex:%d mv:%d misc:%d imb:%d pmb:%d smb:%d d:%c ref:",
1429 h->fenc->i_frame, h->i_frame,
1430 c_type, h->fenc->i_duration,
1431 h->fenc->i_cpb_duration, rc->qpa_rc,
1432 h->stat.frame.i_tex_bits,
1433 h->stat.frame.i_mv_bits,
1434 h->stat.frame.i_misc_bits,
1435 h->stat.frame.i_mb_count_i,
1436 h->stat.frame.i_mb_count_p,
1437 h->stat.frame.i_mb_count_skip,
1441 /* Only write information for reference reordering once. */
1442 int use_old_stats = h->param.rc.b_stat_read && rc->rce->refs > 1;
1443 for( int i = 0; i < (use_old_stats ? rc->rce->refs : h->i_ref0); i++ )
1445 int refcount = use_old_stats ? rc->rce->refcount[i]
1446 : h->param.b_interlaced ? h->stat.frame.i_mb_count_ref[0][i*2]
1447 + h->stat.frame.i_mb_count_ref[0][i*2+1]
1448 : h->stat.frame.i_mb_count_ref[0][i];
1449 if( fprintf( rc->p_stat_file_out, "%d ", refcount ) < 0 )
1453 if( h->sh.weight[0][0].weightfn )
1455 if( fprintf( rc->p_stat_file_out, "w:%"PRId32",%"PRId32",%"PRId32, h->sh.weight[0][0].i_denom, h->sh.weight[0][0].i_scale, h->sh.weight[0][0].i_offset ) < 0 )
1459 if( fprintf( rc->p_stat_file_out, ";\n") < 0 )
1462 /* Don't re-write the data in multi-pass mode. */
1463 if( h->param.rc.b_mb_tree && h->fenc->b_kept_as_ref && !h->param.rc.b_stat_read )
1465 uint8_t i_type = h->sh.i_type;
1466 /* Values are stored as big-endian FIX8.8 */
1467 for( int i = 0; i < h->mb.i_mb_count; i++ )
1468 rc->qp_buffer[0][i] = endian_fix16( h->fenc->f_qp_offset[i]*256.0 );
1469 if( fwrite( &i_type, 1, 1, rc->p_mbtree_stat_file_out ) < 1 )
1471 if( fwrite( rc->qp_buffer[0], sizeof(uint16_t), h->mb.i_mb_count, rc->p_mbtree_stat_file_out ) < h->mb.i_mb_count )
1478 if( h->sh.i_type != SLICE_TYPE_B )
1479 rc->cplxr_sum += bits * qp2qscale( rc->qpa_rc ) / rc->last_rceq;
1482 /* Depends on the fact that B-frame's QP is an offset from the following P-frame's.
1483 * Not perfectly accurate with B-refs, but good enough. */
1484 rc->cplxr_sum += bits * qp2qscale( rc->qpa_rc ) / (rc->last_rceq * fabs( h->param.rc.f_pb_factor ));
1486 rc->cplxr_sum *= rc->cbr_decay;
1487 double frame_duration = (double)h->fenc->i_duration * h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1489 rc->wanted_bits_window += frame_duration * rc->bitrate;
1490 rc->wanted_bits_window *= rc->cbr_decay;
1494 rc->expected_bits_sum += qscale2bits( rc->rce, qp2qscale( rc->rce->new_qp ) );
1496 if( h->mb.b_variable_qp )
1498 if( h->sh.i_type == SLICE_TYPE_B )
1500 rc->bframe_bits += bits;
1501 if( h->fenc->b_last_minigop_bframe )
1503 update_predictor( rc->pred_b_from_p, qp2qscale( rc->qpa_rc ),
1504 h->fref1[h->i_ref1-1]->i_satd, rc->bframe_bits / rc->bframes );
1505 rc->bframe_bits = 0;
1510 *filler = update_vbv( h, bits );
1512 if( h->sps->vui.b_nal_hrd_parameters_present )
1514 if( h->fenc->i_frame == 0 )
1516 // access unit initialises the HRD
1517 h->fenc->hrd_timing.cpb_initial_arrival_time = 0;
1518 rc->initial_cpb_removal_delay = h->initial_cpb_removal_delay;
1519 rc->initial_cpb_removal_delay_offset = h->initial_cpb_removal_delay_offset;
1520 h->fenc->hrd_timing.cpb_removal_time = rc->nrt_first_access_unit = (double)rc->initial_cpb_removal_delay / 90000;
1524 h->fenc->hrd_timing.cpb_removal_time = rc->nrt_first_access_unit + (double)h->fenc->i_cpb_delay *
1525 h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
1527 double cpb_earliest_arrival_time = h->fenc->hrd_timing.cpb_removal_time - (double)rc->initial_cpb_removal_delay / 90000;
1528 if( h->fenc->b_keyframe )
1530 rc->nrt_first_access_unit = h->fenc->hrd_timing.cpb_removal_time;
1531 rc->initial_cpb_removal_delay = h->initial_cpb_removal_delay;
1532 rc->initial_cpb_removal_delay_offset = h->initial_cpb_removal_delay_offset;
1535 cpb_earliest_arrival_time -= (double)rc->initial_cpb_removal_delay_offset / 90000;
1537 if( h->sps->vui.hrd.b_cbr_hrd )
1538 h->fenc->hrd_timing.cpb_initial_arrival_time = rc->previous_cpb_final_arrival_time;
1540 h->fenc->hrd_timing.cpb_initial_arrival_time = X264_MAX( rc->previous_cpb_final_arrival_time, cpb_earliest_arrival_time );
1542 int filler_bits = *filler ? X264_MAX( (FILLER_OVERHEAD - h->param.b_annexb), *filler )*8 : 0;
1544 h->fenc->hrd_timing.cpb_final_arrival_time = rc->previous_cpb_final_arrival_time = h->fenc->hrd_timing.cpb_initial_arrival_time +
1545 (double)(bits + filler_bits) / h->sps->vui.hrd.i_bit_rate_unscaled;
1547 h->fenc->hrd_timing.dpb_output_time = (double)h->fenc->i_dpb_output_delay * h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale +
1548 h->fenc->hrd_timing.cpb_removal_time;
1553 x264_log(h, X264_LOG_ERROR, "ratecontrol_end: stats file could not be written to\n");
1557 /****************************************************************************
1559 ***************************************************************************/
1562 * modify the bitrate curve from pass1 for one frame
1564 static double get_qscale(x264_t *h, ratecontrol_entry_t *rce, double rate_factor, int frame_num)
1566 x264_ratecontrol_t *rcc= h->rc;
1567 x264_zone_t *zone = get_zone( h, frame_num );
1568 double q = pow( rce->blurred_complexity, 1 - rcc->qcompress );
1570 // avoid NaN's in the rc_eq
1571 if( !isfinite(q) || rce->tex_bits + rce->mv_bits == 0 )
1572 q = rcc->last_qscale_for[rce->pict_type];
1577 rcc->last_qscale = q;
1582 if( zone->b_force_qp )
1583 q = qp2qscale( zone->i_qp );
1585 q /= zone->f_bitrate_factor;
1591 static double get_diff_limited_q(x264_t *h, ratecontrol_entry_t *rce, double q)
1593 x264_ratecontrol_t *rcc = h->rc;
1594 const int pict_type = rce->pict_type;
1596 // force I/B quants as a function of P quants
1597 const double last_p_q = rcc->last_qscale_for[SLICE_TYPE_P];
1598 const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
1599 if( pict_type == SLICE_TYPE_I )
1602 double pq = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
1603 double ip_factor = fabs( h->param.rc.f_ip_factor );
1604 /* don't apply ip_factor if the following frame is also I */
1605 if( rcc->accum_p_norm <= 0 )
1607 else if( h->param.rc.f_ip_factor < 0 )
1609 else if( rcc->accum_p_norm >= 1 )
1612 q = rcc->accum_p_norm * pq / ip_factor + (1 - rcc->accum_p_norm) * iq;
1614 else if( pict_type == SLICE_TYPE_B )
1616 if( h->param.rc.f_pb_factor > 0 )
1618 if( !rce->kept_as_ref )
1619 q *= fabs( h->param.rc.f_pb_factor );
1621 else if( pict_type == SLICE_TYPE_P
1622 && rcc->last_non_b_pict_type == SLICE_TYPE_P
1623 && rce->tex_bits == 0 )
1628 /* last qscale / qdiff stuff */
1629 if( rcc->last_non_b_pict_type == pict_type &&
1630 (pict_type!=SLICE_TYPE_I || rcc->last_accum_p_norm < 1) )
1632 double last_q = rcc->last_qscale_for[pict_type];
1633 double max_qscale = last_q * rcc->lstep;
1634 double min_qscale = last_q / rcc->lstep;
1636 if ( q > max_qscale ) q = max_qscale;
1637 else if( q < min_qscale ) q = min_qscale;
1640 rcc->last_qscale_for[pict_type] = q;
1641 if( pict_type != SLICE_TYPE_B )
1642 rcc->last_non_b_pict_type = pict_type;
1643 if( pict_type == SLICE_TYPE_I )
1645 rcc->last_accum_p_norm = rcc->accum_p_norm;
1646 rcc->accum_p_norm = 0;
1647 rcc->accum_p_qp = 0;
1649 if( pict_type == SLICE_TYPE_P )
1651 float mask = 1 - pow( (float)rce->i_count / rcc->nmb, 2 );
1652 rcc->accum_p_qp = mask * (qscale2qp( q ) + rcc->accum_p_qp);
1653 rcc->accum_p_norm = mask * (1 + rcc->accum_p_norm);
1658 static double predict_size( predictor_t *p, double q, double var )
1660 return (p->coeff*var + p->offset) / (q*p->count);
1663 static void update_predictor( predictor_t *p, double q, double var, double bits )
1665 const double range = 1.5;
1668 double old_coeff = p->coeff / p->count;
1669 double new_coeff = bits*q / var;
1670 double new_coeff_clipped = x264_clip3f( new_coeff, old_coeff/range, old_coeff*range );
1671 double new_offset = bits*q - new_coeff_clipped * var;
1672 if( new_offset >= 0 )
1673 new_coeff = new_coeff_clipped;
1676 p->count *= p->decay;
1677 p->coeff *= p->decay;
1678 p->offset *= p->decay;
1680 p->coeff += new_coeff;
1681 p->offset += new_offset;
1684 // update VBV after encoding a frame
1685 static int update_vbv( x264_t *h, int bits )
1689 x264_ratecontrol_t *rcc = h->rc;
1690 x264_ratecontrol_t *rct = h->thread[0]->rc;
1692 if( rcc->last_satd >= h->mb.i_mb_count )
1693 update_predictor( &rct->pred[h->sh.i_type], qp2qscale( rcc->qpa_rc ), rcc->last_satd, bits );
1698 rct->buffer_fill_final -= bits;
1700 if( rct->buffer_fill_final < 0 )
1701 x264_log( h, X264_LOG_WARNING, "VBV underflow (frame %d, %.0f bits)\n", h->i_frame, rct->buffer_fill_final );
1702 rct->buffer_fill_final = X264_MAX( rct->buffer_fill_final, 0 );
1703 rct->buffer_fill_final += rcc->buffer_rate;
1705 if( h->sps->vui.hrd.b_cbr_hrd && rct->buffer_fill_final > rcc->buffer_size )
1707 filler = ceil( (rct->buffer_fill_final - rcc->buffer_size) / 8 );
1708 rct->buffer_fill_final -= X264_MAX( (FILLER_OVERHEAD - h->param.b_annexb), filler ) * 8;
1711 rct->buffer_fill_final = X264_MIN( rct->buffer_fill_final, rcc->buffer_size );
1716 int x264_hrd_fullness( x264_t *h )
1718 x264_ratecontrol_t *rct = h->thread[0]->rc;
1719 double cpb_bits = rct->buffer_fill_final;
1720 double bps = h->sps->vui.hrd.i_bit_rate_unscaled;
1721 double cpb_size = h->sps->vui.hrd.i_cpb_size_unscaled;
1722 double cpb_fullness = 90000.0*cpb_bits/bps;
1724 if( cpb_bits < 0 || cpb_bits > cpb_size )
1726 x264_log( h, X264_LOG_WARNING, "CPB %s: %.0lf bits in a %.0lf-bit buffer\n",
1727 cpb_bits < 0 ? "underflow" : "overflow", cpb_bits, cpb_size );
1730 h->initial_cpb_removal_delay_offset = 90000.0*(cpb_size - cpb_bits)/bps;
1732 return x264_clip3f( cpb_fullness + 0.5, 0, 90000.0*cpb_size/bps ); // just lie if we are in a weird state
1735 // provisionally update VBV according to the planned size of all frames currently in progress
1736 static void update_vbv_plan( x264_t *h, int overhead )
1738 x264_ratecontrol_t *rcc = h->rc;
1739 rcc->buffer_fill = h->thread[0]->rc->buffer_fill_final;
1740 if( h->i_thread_frames > 1 )
1742 int j = h->rc - h->thread[0]->rc;
1743 for( int i = 1; i < h->i_thread_frames; i++ )
1745 x264_t *t = h->thread[ (j+i)%h->i_thread_frames ];
1746 double bits = t->rc->frame_size_planned;
1747 if( !t->b_thread_active )
1749 bits = X264_MAX(bits, t->rc->frame_size_estimated);
1750 rcc->buffer_fill -= bits;
1751 rcc->buffer_fill = X264_MAX( rcc->buffer_fill, 0 );
1752 rcc->buffer_fill += t->rc->buffer_rate;
1753 rcc->buffer_fill = X264_MIN( rcc->buffer_fill, rcc->buffer_size );
1756 rcc->buffer_fill = X264_MIN( rcc->buffer_fill, rcc->buffer_size );
1757 rcc->buffer_fill -= overhead;
1760 // apply VBV constraints and clip qscale to between lmin and lmax
1761 static double clip_qscale( x264_t *h, int pict_type, double q )
1763 x264_ratecontrol_t *rcc = h->rc;
1764 double lmin = rcc->lmin[pict_type];
1765 double lmax = rcc->lmax[pict_type];
1766 if( rcc->rate_factor_max_increment )
1767 lmax = X264_MIN( lmax, qp2qscale( rcc->qp_novbv + rcc->rate_factor_max_increment ) );
1770 /* B-frames are not directly subject to VBV,
1771 * since they are controlled by the P-frames' QPs. */
1773 if( rcc->b_vbv && rcc->last_satd > 0 )
1775 /* Lookahead VBV: raise the quantizer as necessary such that no frames in
1776 * the lookahead overflow and such that the buffer is in a reasonable state
1777 * by the end of the lookahead. */
1778 if( h->param.rc.i_lookahead )
1782 /* Avoid an infinite loop. */
1783 for( int iterations = 0; iterations < 1000 && terminate != 3; iterations++ )
1786 double cur_bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1787 double buffer_fill_cur = rcc->buffer_fill - cur_bits;
1789 double total_duration = 0;
1790 frame_q[0] = h->sh.i_type == SLICE_TYPE_I ? q * h->param.rc.f_ip_factor : q;
1791 frame_q[1] = frame_q[0] * h->param.rc.f_pb_factor;
1792 frame_q[2] = frame_q[0] / h->param.rc.f_ip_factor;
1794 /* Loop over the planned future frames. */
1795 for( int j = 0; buffer_fill_cur >= 0 && buffer_fill_cur <= rcc->buffer_size; j++ )
1797 total_duration += h->fenc->f_planned_cpb_duration[j];
1798 buffer_fill_cur += rcc->vbv_max_rate * h->fenc->f_planned_cpb_duration[j];
1799 int i_type = h->fenc->i_planned_type[j];
1800 int i_satd = h->fenc->i_planned_satd[j];
1801 if( i_type == X264_TYPE_AUTO )
1803 i_type = IS_X264_TYPE_I( i_type ) ? SLICE_TYPE_I : IS_X264_TYPE_B( i_type ) ? SLICE_TYPE_B : SLICE_TYPE_P;
1804 cur_bits = predict_size( &rcc->pred[i_type], frame_q[i_type], i_satd );
1805 buffer_fill_cur -= cur_bits;
1807 /* Try to get to get the buffer at least 50% filled, but don't set an impossible goal. */
1808 target_fill = X264_MIN( rcc->buffer_fill + total_duration * rcc->vbv_max_rate * 0.5, rcc->buffer_size * 0.5 );
1809 if( buffer_fill_cur < target_fill )
1815 /* Try to get the buffer no more than 80% filled, but don't set an impossible goal. */
1816 target_fill = x264_clip3f( rcc->buffer_fill - total_duration * rcc->vbv_max_rate * 0.5, rcc->buffer_size * 0.8, rcc->buffer_size );
1817 if( rcc->b_vbv_min_rate && buffer_fill_cur > target_fill )
1826 /* Fallback to old purely-reactive algorithm: no lookahead. */
1829 if( ( pict_type == SLICE_TYPE_P ||
1830 ( pict_type == SLICE_TYPE_I && rcc->last_non_b_pict_type == SLICE_TYPE_I ) ) &&
1831 rcc->buffer_fill/rcc->buffer_size < 0.5 )
1833 q /= x264_clip3f( 2.0*rcc->buffer_fill/rcc->buffer_size, 0.5, 1.0 );
1836 /* Now a hard threshold to make sure the frame fits in VBV.
1837 * This one is mostly for I-frames. */
1838 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1840 /* For small VBVs, allow the frame to use up the entire VBV. */
1841 double max_fill_factor = h->param.rc.i_vbv_buffer_size >= 5*h->param.rc.i_vbv_max_bitrate / rcc->fps ? 2 : 1;
1842 /* For single-frame VBVs, request that the frame use up the entire VBV. */
1843 double min_fill_factor = rcc->single_frame_vbv ? 1 : 2;
1845 if( bits > rcc->buffer_fill/max_fill_factor )
1846 qf = x264_clip3f( rcc->buffer_fill/(max_fill_factor*bits), 0.2, 1.0 );
1849 if( bits < rcc->buffer_rate/min_fill_factor )
1850 q *= bits*min_fill_factor/rcc->buffer_rate;
1851 q = X264_MAX( q0, q );
1854 /* Apply MinCR restrictions */
1855 double bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1856 if( bits > rcc->frame_size_maximum )
1857 q *= bits / rcc->frame_size_maximum;
1858 bits = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
1860 /* Check B-frame complexity, and use up any bits that would
1861 * overflow before the next P-frame. */
1862 if( h->sh.i_type == SLICE_TYPE_P && !rcc->single_frame_vbv )
1864 int nb = rcc->bframes;
1865 double pbbits = bits;
1866 double bbits = predict_size( rcc->pred_b_from_p, q * h->param.rc.f_pb_factor, rcc->last_satd );
1868 double bframe_cpb_duration = 0;
1869 double minigop_cpb_duration;
1870 for( int i = 0; i < nb; i++ )
1871 bframe_cpb_duration += h->fenc->f_planned_cpb_duration[1+i];
1873 if( bbits * nb > bframe_cpb_duration * rcc->vbv_max_rate )
1875 pbbits += nb * bbits;
1877 minigop_cpb_duration = bframe_cpb_duration + h->fenc->f_planned_cpb_duration[0];
1878 space = rcc->buffer_fill + minigop_cpb_duration*rcc->vbv_max_rate - rcc->buffer_size;
1879 if( pbbits < space )
1881 q *= X264_MAX( pbbits / space, bits / (0.5 * rcc->buffer_size) );
1883 q = X264_MAX( q0-5, q );
1886 if( !rcc->b_vbv_min_rate )
1887 q = X264_MAX( q0, q );
1892 else if( rcc->b_2pass )
1894 double min2 = log( lmin );
1895 double max2 = log( lmax );
1896 q = (log(q) - min2)/(max2-min2) - 0.5;
1897 q = 1.0/(1.0 + exp( -4*q ));
1898 q = q*(max2-min2) + min2;
1902 return x264_clip3f( q, lmin, lmax );
1905 // update qscale for 1 frame based on actual bits used so far
1906 static float rate_estimate_qscale( x264_t *h )
1909 x264_ratecontrol_t *rcc = h->rc;
1910 ratecontrol_entry_t rce;
1911 int pict_type = h->sh.i_type;
1912 int64_t total_bits = 8*(h->stat.i_frame_size[SLICE_TYPE_I]
1913 + h->stat.i_frame_size[SLICE_TYPE_P]
1914 + h->stat.i_frame_size[SLICE_TYPE_B]);
1919 if( pict_type != rce.pict_type )
1921 x264_log( h, X264_LOG_ERROR, "slice=%c but 2pass stats say %c\n",
1922 slice_type_to_char[pict_type], slice_type_to_char[rce.pict_type] );
1926 if( pict_type == SLICE_TYPE_B )
1928 /* B-frames don't have independent ratecontrol, but rather get the
1929 * average QP of the two adjacent P-frames + an offset */
1931 int i0 = IS_X264_TYPE_I(h->fref0[0]->i_type);
1932 int i1 = IS_X264_TYPE_I(h->fref1[0]->i_type);
1933 int dt0 = abs(h->fenc->i_poc - h->fref0[0]->i_poc);
1934 int dt1 = abs(h->fenc->i_poc - h->fref1[0]->i_poc);
1935 float q0 = h->fref0[0]->f_qp_avg_rc;
1936 float q1 = h->fref1[0]->f_qp_avg_rc;
1938 if( h->fref0[0]->i_type == X264_TYPE_BREF )
1939 q0 -= rcc->pb_offset/2;
1940 if( h->fref1[0]->i_type == X264_TYPE_BREF )
1941 q1 -= rcc->pb_offset/2;
1944 q = (q0 + q1) / 2 + rcc->ip_offset;
1950 q = (q0*dt1 + q1*dt0) / (dt0 + dt1);
1952 if( h->fenc->b_kept_as_ref )
1953 q += rcc->pb_offset/2;
1955 q += rcc->pb_offset;
1957 if( rcc->b_2pass && rcc->b_vbv )
1958 rcc->frame_size_planned = qscale2bits( &rce, q );
1960 rcc->frame_size_planned = predict_size( rcc->pred_b_from_p, q, h->fref1[h->i_ref1-1]->i_satd );
1961 h->rc->frame_size_estimated = rcc->frame_size_planned;
1965 rcc->last_satd = x264_rc_analyse_slice( h );
1967 return qp2qscale( q );
1971 double abr_buffer = 2 * rcc->rate_tolerance * rcc->bitrate;
1975 double lmin = rcc->lmin[pict_type];
1976 double lmax = rcc->lmax[pict_type];
1978 int64_t predicted_bits = total_bits;
1979 /* Adjust ABR buffer based on distance to the end of the video. */
1980 if( rcc->num_entries > h->fenc->i_frame )
1981 abr_buffer *= 0.5 * sqrt( rcc->num_entries - h->fenc->i_frame );
1985 if( h->i_thread_frames > 1 )
1987 int j = h->rc - h->thread[0]->rc;
1988 for( int i = 1; i < h->i_thread_frames; i++ )
1990 x264_t *t = h->thread[ (j+i)%h->i_thread_frames ];
1991 double bits = t->rc->frame_size_planned;
1992 if( !t->b_thread_active )
1994 bits = X264_MAX(bits, t->rc->frame_size_estimated);
1995 predicted_bits += (int64_t)bits;
2001 if( h->fenc->i_frame < h->i_thread_frames )
2002 predicted_bits += (int64_t)h->fenc->i_frame * rcc->bitrate / rcc->fps;
2004 predicted_bits += (int64_t)(h->i_thread_frames - 1) * rcc->bitrate / rcc->fps;
2007 diff = predicted_bits - (int64_t)rce.expected_bits;
2009 q /= x264_clip3f((double)(abr_buffer - diff) / abr_buffer, .5, 2);
2010 if( ((h->fenc->i_frame + 1 - h->i_thread_frames) >= rcc->fps) &&
2011 (rcc->expected_bits_sum > 0))
2013 /* Adjust quant based on the difference between
2014 * achieved and expected bitrate so far */
2015 double cur_time = (double)h->fenc->i_frame / rcc->num_entries;
2016 double w = x264_clip3f( cur_time*100, 0.0, 1.0 );
2017 q *= pow( (double)total_bits / rcc->expected_bits_sum, w );
2021 /* Do not overflow vbv */
2022 double expected_size = qscale2bits( &rce, q );
2023 double expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
2024 double expected_fullness = rce.expected_vbv / rcc->buffer_size;
2025 double qmax = q*(2 - expected_fullness);
2026 double size_constraint = 1 + expected_fullness;
2027 qmax = X264_MAX( qmax, rce.new_qscale );
2028 if( expected_fullness < .05 )
2030 qmax = X264_MIN(qmax, lmax);
2031 while( ((expected_vbv < rce.expected_vbv/size_constraint) && (q < qmax)) ||
2032 ((expected_vbv < 0) && (q < lmax)))
2035 expected_size = qscale2bits(&rce, q);
2036 expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
2038 rcc->last_satd = x264_rc_analyse_slice( h );
2040 q = x264_clip3f( q, lmin, lmax );
2042 else /* 1pass ABR */
2044 /* Calculate the quantizer which would have produced the desired
2045 * average bitrate if it had been applied to all frames so far.
2046 * Then modulate that quant based on the current frame's complexity
2047 * relative to the average complexity so far (using the 2pass RCEQ).
2048 * Then bias the quant up or down if total size so far was far from
2050 * Result: Depending on the value of rate_tolerance, there is a
2051 * tradeoff between quality and bitrate precision. But at large
2052 * tolerances, the bit distribution approaches that of 2pass. */
2054 double wanted_bits, overflow = 1;
2056 rcc->last_satd = x264_rc_analyse_slice( h );
2057 rcc->short_term_cplxsum *= 0.5;
2058 rcc->short_term_cplxcount *= 0.5;
2059 rcc->short_term_cplxsum += rcc->last_satd;
2060 rcc->short_term_cplxcount ++;
2062 rce.tex_bits = rcc->last_satd;
2063 rce.blurred_complexity = rcc->short_term_cplxsum / rcc->short_term_cplxcount;
2065 rce.p_count = rcc->nmb;
2069 rce.pict_type = pict_type;
2071 if( h->param.rc.i_rc_method == X264_RC_CRF )
2073 q = get_qscale( h, &rce, rcc->rate_factor_constant, h->fenc->i_frame );
2077 int i_frame_done = h->fenc->i_frame + 1 - h->i_thread_frames;
2078 double i_time_done = i_frame_done / rcc->fps;
2079 if( h->param.b_vfr_input )
2080 i_time_done = ((double)(h->fenc->i_reordered_pts - h->first_pts)) * h->param.i_timebase_num / h->param.i_timebase_den;
2082 q = get_qscale( h, &rce, rcc->wanted_bits_window / rcc->cplxr_sum, h->fenc->i_frame );
2084 /* ABR code can potentially be counterproductive in CBR, so just don't bother.
2085 * Don't run it if the frame complexity is zero either. */
2086 if( !rcc->b_vbv_min_rate && rcc->last_satd )
2088 // FIXME is it simpler to keep track of wanted_bits in ratecontrol_end?
2089 wanted_bits = i_time_done * rcc->bitrate;
2090 if( wanted_bits > 0 )
2092 abr_buffer *= X264_MAX( 1, sqrt(i_time_done) );
2093 overflow = x264_clip3f( 1.0 + (total_bits - wanted_bits) / abr_buffer, .5, 2 );
2099 if( pict_type == SLICE_TYPE_I && h->param.i_keyint_max > 1
2100 /* should test _next_ pict type, but that isn't decided yet */
2101 && rcc->last_non_b_pict_type != SLICE_TYPE_I )
2103 q = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm );
2104 q /= fabs( h->param.rc.f_ip_factor );
2106 else if( h->i_frame > 0 )
2108 /* Asymmetric clipping, because symmetric would prevent
2109 * overflow control in areas of rapidly oscillating complexity */
2110 double lmin = rcc->last_qscale_for[pict_type] / rcc->lstep;
2111 double lmax = rcc->last_qscale_for[pict_type] * rcc->lstep;
2112 if( overflow > 1.1 && h->i_frame > 3 )
2114 else if( overflow < 0.9 )
2117 q = x264_clip3f(q, lmin, lmax);
2119 else if( h->param.rc.i_rc_method == X264_RC_CRF && rcc->qcompress != 1 )
2121 q = qp2qscale( ABR_INIT_QP ) / fabs( h->param.rc.f_ip_factor );
2123 rcc->qp_novbv = qscale2qp( q );
2125 //FIXME use get_diff_limited_q() ?
2126 q = clip_qscale( h, pict_type, q );
2129 rcc->last_qscale_for[pict_type] =
2130 rcc->last_qscale = q;
2132 if( !(rcc->b_2pass && !rcc->b_vbv) && h->fenc->i_frame == 0 )
2133 rcc->last_qscale_for[SLICE_TYPE_P] = q * fabs( h->param.rc.f_ip_factor );
2135 if( rcc->b_2pass && rcc->b_vbv )
2136 rcc->frame_size_planned = qscale2bits(&rce, q);
2138 rcc->frame_size_planned = predict_size( &rcc->pred[h->sh.i_type], q, rcc->last_satd );
2140 /* Always use up the whole VBV in this case. */
2141 if( rcc->single_frame_vbv )
2142 rcc->frame_size_planned = rcc->buffer_rate;
2143 h->rc->frame_size_estimated = rcc->frame_size_planned;
2148 void x264_threads_normalize_predictors( x264_t *h )
2150 double totalsize = 0;
2151 for( int i = 0; i < h->param.i_threads; i++ )
2152 totalsize += h->thread[i]->rc->slice_size_planned;
2153 double factor = h->rc->frame_size_planned / totalsize;
2154 for( int i = 0; i < h->param.i_threads; i++ )
2155 h->thread[i]->rc->slice_size_planned *= factor;
2158 void x264_threads_distribute_ratecontrol( x264_t *h )
2161 x264_ratecontrol_t *rc = h->rc;
2163 /* Initialize row predictors */
2164 if( h->i_frame == 0 )
2165 for( int i = 0; i < h->param.i_threads; i++ )
2167 x264_ratecontrol_t *t = h->thread[i]->rc;
2168 memcpy( t->row_preds, rc->row_preds, sizeof(rc->row_preds) );
2171 for( int i = 0; i < h->param.i_threads; i++ )
2173 x264_t *t = h->thread[i];
2174 memcpy( t->rc, rc, offsetof(x264_ratecontrol_t, row_pred) );
2175 t->rc->row_pred = &t->rc->row_preds[h->sh.i_type];
2176 /* Calculate the planned slice size. */
2177 if( rc->b_vbv && rc->frame_size_planned )
2180 for( row = t->i_threadslice_start; row < t->i_threadslice_end; row++ )
2181 size += h->fdec->i_row_satd[row];
2182 t->rc->slice_size_planned = predict_size( &rc->pred[h->sh.i_type + (i+1)*5], rc->qpm, size );
2185 t->rc->slice_size_planned = 0;
2187 if( rc->b_vbv && rc->frame_size_planned )
2189 x264_threads_normalize_predictors( h );
2191 if( rc->single_frame_vbv )
2193 /* Compensate for our max frame error threshold: give more bits (proportionally) to smaller slices. */
2194 for( int i = 0; i < h->param.i_threads; i++ )
2196 x264_t *t = h->thread[i];
2197 t->rc->max_frame_error = X264_MAX( 0.05, 1.0 / (t->i_threadslice_end - t->i_threadslice_start) );
2198 t->rc->slice_size_planned += 2 * t->rc->max_frame_error * rc->frame_size_planned;
2200 x264_threads_normalize_predictors( h );
2203 for( int i = 0; i < h->param.i_threads; i++ )
2204 h->thread[i]->rc->frame_size_estimated = h->thread[i]->rc->slice_size_planned;
2208 void x264_threads_merge_ratecontrol( x264_t *h )
2210 x264_ratecontrol_t *rc = h->rc;
2213 for( int i = 0; i < h->param.i_threads; i++ )
2215 x264_t *t = h->thread[i];
2216 x264_ratecontrol_t *rct = h->thread[i]->rc;
2217 if( h->param.rc.i_vbv_buffer_size )
2220 for( int row = t->i_threadslice_start; row < t->i_threadslice_end; row++ )
2221 size += h->fdec->i_row_satd[row];
2222 int bits = t->stat.frame.i_mv_bits + t->stat.frame.i_tex_bits + t->stat.frame.i_misc_bits;
2223 int mb_count = (t->i_threadslice_end - t->i_threadslice_start) * h->sps->i_mb_width;
2224 update_predictor( &rc->pred[h->sh.i_type+5*i], qp2qscale( rct->qpa_rc/mb_count ), size, bits );
2228 rc->qpa_rc += rct->qpa_rc;
2229 rc->qpa_aq += rct->qpa_aq;
2233 void x264_thread_sync_ratecontrol( x264_t *cur, x264_t *prev, x264_t *next )
2237 #define COPY(var) memcpy(&cur->rc->var, &prev->rc->var, sizeof(cur->rc->var))
2238 /* these vars are updated in x264_ratecontrol_start()
2239 * so copy them from the context that most recently started (prev)
2240 * to the context that's about to start (cur). */
2245 COPY(last_qscale_for);
2246 COPY(last_non_b_pict_type);
2247 COPY(short_term_cplxsum);
2248 COPY(short_term_cplxcount);
2252 /* these vars can be updated by x264_ratecontrol_init_reconfigurable */
2255 COPY(single_frame_vbv);
2257 COPY(b_vbv_min_rate);
2258 COPY(rate_factor_constant);
2264 #define COPY(var) next->rc->var = cur->rc->var
2265 /* these vars are updated in x264_ratecontrol_end()
2266 * so copy them from the context that most recently ended (cur)
2267 * to the context that's about to end (next) */
2269 COPY(expected_bits_sum);
2270 COPY(wanted_bits_window);
2272 COPY(initial_cpb_removal_delay);
2273 COPY(initial_cpb_removal_delay_offset);
2274 COPY(nrt_first_access_unit);
2275 COPY(previous_cpb_final_arrival_time);
2278 //FIXME row_preds[] (not strictly necessary, but would improve prediction)
2279 /* the rest of the variables are either constant or thread-local */
2282 static int find_underflow( x264_t *h, double *fills, int *t0, int *t1, int over )
2284 /* find an interval ending on an overflow or underflow (depending on whether
2285 * we're adding or removing bits), and starting on the earliest frame that
2286 * can influence the buffer fill of that end frame. */
2287 x264_ratecontrol_t *rcc = h->rc;
2288 const double buffer_min = (over ? .1 : .1) * rcc->buffer_size;
2289 const double buffer_max = .9 * rcc->buffer_size;
2290 double fill = fills[*t0-1];
2291 double parity = over ? 1. : -1.;
2292 int start = -1, end = -1;
2293 for( int i = *t0; i < rcc->num_entries; i++ )
2295 fill += (rcc->entry[i].i_cpb_duration * rcc->vbv_max_rate * h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale -
2296 qscale2bits( &rcc->entry[i], rcc->entry[i].new_qscale )) * parity;
2297 fill = x264_clip3f(fill, 0, rcc->buffer_size);
2299 if( fill <= buffer_min || i == 0 )
2305 else if( fill >= buffer_max && start >= 0 )
2310 return start >= 0 && end >= 0;
2313 static int fix_underflow( x264_t *h, int t0, int t1, double adjustment, double qscale_min, double qscale_max)
2315 x264_ratecontrol_t *rcc = h->rc;
2316 double qscale_orig, qscale_new;
2320 for( int i = t0; i <= t1; i++ )
2322 qscale_orig = rcc->entry[i].new_qscale;
2323 qscale_orig = x264_clip3f( qscale_orig, qscale_min, qscale_max );
2324 qscale_new = qscale_orig * adjustment;
2325 qscale_new = x264_clip3f( qscale_new, qscale_min, qscale_max );
2326 rcc->entry[i].new_qscale = qscale_new;
2327 adjusted = adjusted || (qscale_new != qscale_orig);
2332 static double count_expected_bits( x264_t *h )
2334 x264_ratecontrol_t *rcc = h->rc;
2335 double expected_bits = 0;
2336 for( int i = 0; i < rcc->num_entries; i++ )
2338 ratecontrol_entry_t *rce = &rcc->entry[i];
2339 rce->expected_bits = expected_bits;
2340 expected_bits += qscale2bits( rce, rce->new_qscale );
2342 return expected_bits;
2345 static int vbv_pass2( x264_t *h, double all_available_bits )
2347 /* for each interval of buffer_full .. underflow, uniformly increase the qp of all
2348 * frames in the interval until either buffer is full at some intermediate frame or the
2349 * last frame in the interval no longer underflows. Recompute intervals and repeat.
2350 * Then do the converse to put bits back into overflow areas until target size is met */
2352 x264_ratecontrol_t *rcc = h->rc;
2354 double expected_bits = 0;
2356 double prev_bits = 0;
2358 double qscale_min = qp2qscale( h->param.rc.i_qp_min );
2359 double qscale_max = qp2qscale( h->param.rc.i_qp_max );
2361 int adj_min, adj_max;
2362 CHECKED_MALLOC( fills, (rcc->num_entries+1)*sizeof(double) );
2366 /* adjust overall stream size */
2370 prev_bits = expected_bits;
2373 { /* not first iteration */
2374 adjustment = X264_MAX(X264_MIN(expected_bits / all_available_bits, 0.999), 0.9);
2375 fills[-1] = rcc->buffer_size * h->param.rc.f_vbv_buffer_init;
2379 while(adj_min && find_underflow( h, fills, &t0, &t1, 1 ))
2381 adj_min = fix_underflow( h, t0, t1, adjustment, qscale_min, qscale_max );
2386 fills[-1] = rcc->buffer_size * (1. - h->param.rc.f_vbv_buffer_init);
2388 /* fix underflows -- should be done after overflow, as we'd better undersize target than underflowing VBV */
2390 while( adj_max && find_underflow( h, fills, &t0, &t1, 0 ) )
2391 adj_max = fix_underflow( h, t0, t1, 1.001, qscale_min, qscale_max );
2393 expected_bits = count_expected_bits( h );
2394 } while( (expected_bits < .995*all_available_bits) && ((int64_t)(expected_bits+.5) > (int64_t)(prev_bits+.5)) );
2397 x264_log( h, X264_LOG_WARNING, "vbv-maxrate issue, qpmax or vbv-maxrate too low\n");
2399 /* store expected vbv filling values for tracking when encoding */
2400 for( int i = 0; i < rcc->num_entries; i++ )
2401 rcc->entry[i].expected_vbv = rcc->buffer_size - fills[i];
2403 x264_free( fills-1 );
2409 static int init_pass2( x264_t *h )
2411 x264_ratecontrol_t *rcc = h->rc;
2412 uint64_t all_const_bits = 0;
2413 double duration = 0;
2414 for( int i = 0; i < rcc->num_entries; i++ )
2415 duration += rcc->entry[i].i_duration;
2416 duration *= (double)h->sps->vui.i_num_units_in_tick / h->sps->vui.i_time_scale;
2417 uint64_t all_available_bits = h->param.rc.i_bitrate * 1000. * duration;
2418 double rate_factor, step_mult;
2419 double qblur = h->param.rc.f_qblur;
2420 double cplxblur = h->param.rc.f_complexity_blur;
2421 const int filter_size = (int)(qblur*4) | 1;
2422 double expected_bits;
2423 double *qscale, *blurred_qscale;
2425 /* find total/average complexity & const_bits */
2426 for( int i = 0; i < rcc->num_entries; i++ )
2428 ratecontrol_entry_t *rce = &rcc->entry[i];
2429 all_const_bits += rce->misc_bits;
2432 if( all_available_bits < all_const_bits)
2434 x264_log( h, X264_LOG_ERROR, "requested bitrate is too low. estimated minimum is %d kbps\n",
2435 (int)(all_const_bits * rcc->fps / (rcc->num_entries * 1000.)) );
2439 /* Blur complexities, to reduce local fluctuation of QP.
2440 * We don't blur the QPs directly, because then one very simple frame
2441 * could drag down the QP of a nearby complex frame and give it more
2442 * bits than intended. */
2443 for( int i = 0; i < rcc->num_entries; i++ )
2445 ratecontrol_entry_t *rce = &rcc->entry[i];
2446 double weight_sum = 0;
2447 double cplx_sum = 0;
2448 double weight = 1.0;
2449 double gaussian_weight;
2450 /* weighted average of cplx of future frames */
2451 for( int j = 1; j < cplxblur*2 && j < rcc->num_entries-i; j++ )
2453 ratecontrol_entry_t *rcj = &rcc->entry[i+j];
2454 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2455 if( weight < .0001 )
2457 gaussian_weight = weight * exp( -j*j/200.0 );
2458 weight_sum += gaussian_weight;
2459 cplx_sum += gaussian_weight * (qscale2bits(rcj, 1) - rcj->misc_bits);
2461 /* weighted average of cplx of past frames */
2463 for( int j = 0; j <= cplxblur*2 && j <= i; j++ )
2465 ratecontrol_entry_t *rcj = &rcc->entry[i-j];
2466 gaussian_weight = weight * exp( -j*j/200.0 );
2467 weight_sum += gaussian_weight;
2468 cplx_sum += gaussian_weight * (qscale2bits( rcj, 1 ) - rcj->misc_bits);
2469 weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 );
2470 if( weight < .0001 )
2473 rce->blurred_complexity = cplx_sum / weight_sum;
2476 CHECKED_MALLOC( qscale, sizeof(double)*rcc->num_entries );
2477 if( filter_size > 1 )
2478 CHECKED_MALLOC( blurred_qscale, sizeof(double)*rcc->num_entries );
2480 blurred_qscale = qscale;
2482 /* Search for a factor which, when multiplied by the RCEQ values from
2483 * each frame, adds up to the desired total size.
2484 * There is no exact closed-form solution because of VBV constraints and
2485 * because qscale2bits is not invertible, but we can start with the simple
2486 * approximation of scaling the 1st pass by the ratio of bitrates.
2487 * The search range is probably overkill, but speed doesn't matter here. */
2490 for( int i = 0; i < rcc->num_entries; i++ )
2492 double q = get_qscale(h, &rcc->entry[i], 1.0, i);
2493 expected_bits += qscale2bits(&rcc->entry[i], q);
2494 rcc->last_qscale_for[rcc->entry[i].pict_type] = q;
2496 step_mult = all_available_bits / expected_bits;
2499 for( double step = 1E4 * step_mult; step > 1E-7 * step_mult; step *= 0.5)
2502 rate_factor += step;
2504 rcc->last_non_b_pict_type = -1;
2505 rcc->last_accum_p_norm = 1;
2506 rcc->accum_p_norm = 0;
2509 for( int i = 0; i < rcc->num_entries; i++ )
2511 qscale[i] = get_qscale( h, &rcc->entry[i], rate_factor, i );
2512 rcc->last_qscale_for[rcc->entry[i].pict_type] = qscale[i];
2515 /* fixed I/B qscale relative to P */
2516 for( int i = rcc->num_entries-1; i >= 0; i-- )
2518 qscale[i] = get_diff_limited_q( h, &rcc->entry[i], qscale[i] );
2519 assert(qscale[i] >= 0);
2523 if( filter_size > 1 )
2525 assert( filter_size%2 == 1 );
2526 for( int i = 0; i < rcc->num_entries; i++ )
2528 ratecontrol_entry_t *rce = &rcc->entry[i];
2529 double q = 0.0, sum = 0.0;
2531 for( int j = 0; j < filter_size; j++ )
2533 int index = i+j-filter_size/2;
2535 double coeff = qblur==0 ? 1.0 : exp( -d*d/(qblur*qblur) );
2536 if( index < 0 || index >= rcc->num_entries )
2538 if( rce->pict_type != rcc->entry[index].pict_type )
2540 q += qscale[index] * coeff;
2543 blurred_qscale[i] = q/sum;
2547 /* find expected bits */
2548 for( int i = 0; i < rcc->num_entries; i++ )
2550 ratecontrol_entry_t *rce = &rcc->entry[i];
2551 rce->new_qscale = clip_qscale( h, rce->pict_type, blurred_qscale[i] );
2552 assert(rce->new_qscale >= 0);
2553 expected_bits += qscale2bits( rce, rce->new_qscale );
2556 if( expected_bits > all_available_bits )
2557 rate_factor -= step;
2560 x264_free( qscale );
2561 if( filter_size > 1 )
2562 x264_free( blurred_qscale );
2565 if( vbv_pass2( h, all_available_bits ) )
2567 expected_bits = count_expected_bits( h );
2569 if( fabs( expected_bits/all_available_bits - 1.0 ) > 0.01 )
2572 for( int i = 0; i < rcc->num_entries; i++ )
2573 avgq += rcc->entry[i].new_qscale;
2574 avgq = qscale2qp( avgq / rcc->num_entries );
2576 if( expected_bits > all_available_bits || !rcc->b_vbv )
2577 x264_log( h, X264_LOG_WARNING, "Error: 2pass curve failed to converge\n" );
2578 x264_log( h, X264_LOG_WARNING, "target: %.2f kbit/s, expected: %.2f kbit/s, avg QP: %.4f\n",
2579 (float)h->param.rc.i_bitrate,
2580 expected_bits * rcc->fps / (rcc->num_entries * 1000.),
2582 if( expected_bits < all_available_bits && avgq < h->param.rc.i_qp_min + 2 )
2584 if( h->param.rc.i_qp_min > 0 )
2585 x264_log( h, X264_LOG_WARNING, "try reducing target bitrate or reducing qp_min (currently %d)\n", h->param.rc.i_qp_min );
2587 x264_log( h, X264_LOG_WARNING, "try reducing target bitrate\n" );
2589 else if( expected_bits > all_available_bits && avgq > h->param.rc.i_qp_max - 2 )
2591 if( h->param.rc.i_qp_max < 51 )
2592 x264_log( h, X264_LOG_WARNING, "try increasing target bitrate or increasing qp_max (currently %d)\n", h->param.rc.i_qp_max );
2594 x264_log( h, X264_LOG_WARNING, "try increasing target bitrate\n");
2596 else if( !(rcc->b_2pass && rcc->b_vbv) )
2597 x264_log( h, X264_LOG_WARNING, "internal error\n" );